fix: detect statusline width from controlling terminal

[linziyanleo]

Summary

• When the statusline process's stdout is piped to Claude Code, process.stdout.columns is undefined and terminal width detection fails
• Falls back to a hardcoded default (120) even when the real terminal is narrower or wider, causing broken wrapping or wasted space
• Adds /dev/tty fallback: opens the controlling terminal directly to query its column count when stdout/stderr streams and COLUMNS env are all unavailable
• Skipped on Windows (no /dev/tty) and when CLAUDE_HUD_DISABLE_TTY_WIDTH=1 is set (for testing)

Test plan

• Updated tests/terminal.test.js to set CLAUDE_HUD_DISABLE_TTY_WIDTH=1 in the test harness so unit tests don't accidentally read the real terminal
• Full test suite passes (508/508)
• Manual verification: statusline correctly detects terminal width in Ghostty, iTerm2, and Terminal.app when stdout is piped

[jarrodwatts]

This targets a real piped-stdout width problem, but I am holding it because it adds synchronous /dev/tty probing to the hot statusline path. The HUD can run every few hundred milliseconds, so this needs more proof before it belongs in core width detection.

Required fixes:

• Add a positive mocked test for the /dev/tty fallback and a failure/unavailable-tty test.
• Demonstrate or avoid repeated open/destroy cost in the hot path. A lightweight cache or setup-level width propagation may be better if repeated probing is measurable.
• Confirm Node and Bun behavior for the tty.WriteStream lifecycle, including close/destroy behavior after openSync.

Security review found no token/network issue. The hold is runtime risk and testability.

[jarrodwatts]

Thanks for adding the cache and mocked tty-probe coverage. I am leaving this open for a maintainer width-strategy decision after the forceMaxWidth merge. The remaining question is whether core width detection should probe /dev/tty in the statusline process at all, even with per-process caching.\n\nNo credential or network issue was found. The open risk is runtime behavior in the hot statusline path, plus overlap with the setup-level COLUMNS work in #513 and the broader row-budget work in #511.